Evaluation Methods for Active Human-Guided Neuroevolution in Games (2012)
Machine learning (ML) games such as NERO incorporate human-guided ML methods such as real time neuroevolution (NE) as an integral part of the gameplay, i.e. by allowing the player to train teams of autonomous agents to compete with those trained by others. In order to improve human-guided ML, a way to systematically compare and validate new such methods is needed. To address this problem, this paper describes the results of a human subject study comparing human-guided ML methods and an online tournament validating them at scale. Additionally, this paper describes ongoing work to extend human-guided NE methods through active advice, examples and shaping and to combine these modalities into a more flexible and powerful overall system for agents in games.
In 2012 AAAI Fall Symposium on Robots Learning Interactively from Human Teachers (RLIHT), November 2012.

Leif Johnson leif [at] cs utexas edu
Igor V. Karpov Masters Alumni ikarpov [at] gmail com
Risto Miikkulainen Faculty risto [at] cs utexas edu
Vinod Valsalam Ph.D. Alumni vkv [at] alumni utexas net